A co-evolutionary genetic programming framework for interpretable market-adaptive factor generation in stock selection

This paper proposes a co-evolutionary genetic programming framework for interpretable and market-adaptive factor generation in stock selection. The framework integrates Hidden Markov Models with multiple regime-specific genetic programming components to capture market-level and industry-level state variation. Each component evolves symbolic factor components associated with a specific regime branch, and these components are aggregated through a probabilistic market-industry gating structure to form a composite stock-ranking signal. To improve coordination among components, we introduce a Best-L marginal feedback mechanism that rewards candidate components according to their marginal contribution to the final composite factor. The framework is evaluated using daily data from historical CSI 300 constituent stocks from 2017 to 2025 under a fixed chronological training, validation, and test protocol. Under the main daily rebalanced Top-20 stock selection strategy, the proposed CEGP strategy achieves an annualized return of 34.02% and a Sharpe ratio of 1.35 in the held-out test period, outperforming the all-stock benchmark, XGBoost-Rank, PPO-Ranking, and a plain genetic programming baseline. Additional sorting tests, transaction-cost analysis, feedback-sensitivity tests, alternative portfolio construction settings, and CSI 800 universe robustness results further support the effectiveness of the proposed framework. The findings suggest that structurally guided co-evolution can generate interpretable and adaptive symbolic factors for financial decision support in non-stationary markets.